The precise motion estimation of automatic vehicles using visual sensors is very important since on one hand it contributes to the localization of the vehicles on Automatic Driving 3D maps contributing to the relation between the reality of the roads, route planning and the control of the vehicle; on the other hand, visual sensors provide good readings where other sensors may fail.
Structure from Motion provides an approach to motion estimation using visual sensors. In principle, the Structure from Motion approach requires the estimation of up to 6 parameters to perform the estimation of the autonomous vehicle movement, wherein 3 parameters are required to estimate for the rotation around the three coordinate axes of 3D spaces and 3 parameters for the translation along the coordinate axes. In practice, only 5 parameters may be sufficient because the translation can be estimated only with scaling factor.
Nevertheless, the estimation of five parameters may still be too expensive for autonomous vehicles from both the computational point of view and the estimation precision point of view.